Dr. Paul Baggenstoss is working at Fraunhofer FKIE in Wachtberg, Germany. He is a visiting scientist from the Naval Undersea Warfare Center (NUWC) in Newport, Rhode Island, USA, where he works in the field of classification and automation algorithms for underwater acoustic signals. He has a keen interest in the theory of classification and probability density function (PDF) estimation.
Classical decision theory (Bayesian probabilistic classifier) requires the estimation of PDF's using a common feature set. A feature set that is general enough to represent the characteristics of all classes must be used. This prevents the classifier designer from using special signal processing methods individually for each class. To solve this limitation, I present a rigorous theoretical framework for the use of class-dependent features in Bayesian classifier. The newest extension of this theory is the multi-resolution hidden Markov model (MR-HMM) which integrates several signal processing approaches, each with different time resolution, into a single probabilistic model. In my talk, I will cover the following topics:
* PDF estimation and dimensionality.
* Generative vs. discriminative classifiers.
* Classifying with class-dependent features.
* Probabilistic Multi-resolution signal analysis (MR-HMM).